David Frantz

Characterization of forest disturbance using data-blending techniques applied to MODIS and Landsat time series in an Australian Savanna

Spatio-temporal information on forest loss is essential for a wide range of applications, among them international agreements like the Kyoto Protocol or REDD, or the enforcement of land regulation jurisdiction. Despite remote sensing being the only feasible means of monitoring forest change at regional or greater scales, there is currently no remote sensor that meets the demand of monitoring forest change at landscape level with guaranteed high temporal frequency. 

An alternative to relying on a single inadequate dataset is to exploit the complementary attributes of two datasets by using data fusion techniques. The Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) was specifically designed for blending Landsat and MODIS surface reflectance for applications that have demands for high resolution in both space and time. 

To test the feasibility of fused data for disturbance detection, we chose a test site in an Australian savanna where anthropogenic clearing is commonplace. It was investigated whether the generated dataset could supplement information regarding forest clearing, i.e. an improvement in timing.

An operational scheme for deriving Nadir BRDF adjusted reflectance (NBAR) was utilized to radiometrically standardize all available Landsat TM data for a three year period (2007-2009). Landsat NBAR data and equivalently standardized 16-day MODIS NBAR data were used to generate a dense synthetic time series using STARFM. The quality of the derived images was assessed by comparing them with independent Landsat observations from the nearest date. Overall prediction quality was found to be good (0.84 < R² < 0.97), while indicating that there might be some problems in detecting forest loss at sub-MODIS scale. 

Forest loss of the woody part of the study area was detected by applying a time series based disturbance detection approach employing the Disturbance Index. Both the original Landsat time series and the synthetic time series, as well as a combined hybrid approach were used to identify timing and extent of disturbances. The identified clearings were validated by a well established reference dataset that is regularly produced by Queensland's Statewide Landcover and Trees Study (SLATS).

A comparison of the results derived by the original and synthetic time series generally showed a high degree of agreement in spatial terms (i.e. disturbance extent) and was less pronounced with small clearings. The results of the hybrid detection indeed indicated a temporal improvement in disturbance timing relative to the discrete Landsat data.